Qwen3.8-2.4T-A95B review, pricing and verdict
The first Qwen-Max-class flagship Alibaba has open-weighted: a 2.4T-parameter agentic reasoning model succeeding Qwen3.7-Max.
- Best for
- Self-hosted agentic coding pipelines
- Not for
- Native image or video understanding
Should you use Qwen3.8-2.4T-A95B?
Alibaba open-weighted this checkpoint in August 2026, shortly after Qwen3.8-Max's API debut, extending context to roughly 1 million tokens with FP8, GGUF and NVFP4 builds available the same day. It replaces the prior generation as Alibaba's flagship reasoning and agentic-coding model for teams that need self-hosting or multi-cloud deployment instead of a single vendor API.
Best for
- Self-hosted agentic coding pipelines Open weights in FP8/GGUF/NVFP4 plus 95B active-parameter MoE routing keep inference cost tied to a fraction of the 2.4T total.
- Long-context document or codebase analysis Context extends to roughly 1,010,000 tokens from a 262,144-token native window.
- Cost-sensitive agent workloads vs closed frontier APIs $2.00/$6.00 per 1M tokens undercuts Qwen3.7-Max and most proprietary flagship pricing.
Not for
- Native image or video understanding This checkpoint is text-only; Alibaba's multimodal work sits in the separate Qwen3-VL model line.
- Compliance review requiring a published system card No safety system card or training data cutoff date has been published as of August 2026.
Can you use Qwen3.8-2.4T-A95B?
- Alibaba Cloud Model Studio: yes (Pay as you go (a one-off trial of 1M free tokens for 90 days on the Singapore endpoint only)).
- Free plan: not published.
- Open weights: released (Custom licence).
Checked 30 Oct against Alibaba Cloud's own pages.
Qwen3.8-2.4T-A95B pricing
- Input, per 1M tokens
- $2.00/1M
- Output, per 1M tokens
- $6.00/1M
- Cached input, per 1M tokens
- $0.25/1M
- Pricing model
- per-token
- Free tier
- No
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.060 | $0.0060 | $0.066 |
| Support reply | $0.0040 | $0.0018 | $0.0058 |
| One coding agent run | $0.400 | $0.120 | $0.520 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Cost examples
- Summarize 100K-token research paper
- $0.21
- Daily coding agent (1M tokens in / 200K out)
- $3.20
- Customer support chat (2K in / 500 out per turn, 1000 turns)
- $7.00
QwenCloud's International (Singapore) endpoint prices Qwen3.8-2.4T-A95B at $2.00 per 1M input tokens and $6.00 per 1M output tokens, with cached input at $0.25 per 1M (an 8x discount). The Mainland China (Beijing) endpoint runs an estimated 60-70% cheaper for the same model. OpenRouter's routed pricing has floated slightly lower, around $1.80/$5.40 per 1M, depending on which upstream host serves the request. No separate batch-API discount has been published for this model.
Self-hosting requirements
- Self-hosted
- Yes
- Recommended GPU
- NVIDIA GB300 NVL72 (reference deployment per NVIDIA technical blog); multi-GPU/multi-node required at any precision given 2.4T total parameters
What Qwen3.8-2.4T-A95B can do
Qwen3.8-2.4T-A95B is Alibaba Cloud's mixture-of-experts flagship, activating 95 billion of its 2.4 trillion parameters per token. It is the first Qwen-Max-class model Alibaba has released as open weights, aimed at agentic coding and long-horizon reasoning workloads that previously required a closed API.
What goes in and what comes out
Reads
- text
- tool-calls
Writes
- text
- tool-calls
Capabilities
- Function callingSupported
- Structured outputSupported
- Tool useSupported
- Audio inputNot supported
- Audio outputNot supported
- Code executionNot supported
- Video inputNot supported
- VisionNot supported
- Web browsingNot supported
- Reasoning modes
- low, high, xhigh
Key features
Configurable Reasoning Effort
Switches between low, high and xhigh reasoning modes per request, trading inference cost for depth on hard reasoning and coding tasks.
1M-Class Context Window
Native 262,144-token context extends to roughly 1,010,000 tokens, enough for large codebases or multi-session agent transcripts in one call.
Open Weights in Four Formats
Released on Hugging Face as full safetensors, FP8, GGUF (via Unsloth) and NVFP4 (via RadixArk) checkpoints for self-hosted deployment.
Fine-Grained MoE Routing
512 total experts with 10 routed plus 1 shared expert active per token, keeping inference cost tied to the 95B active count rather than the full 2.4T parameters.
Day-0 Inference Stack Support
Four independent serving stacks added same-day support rather than waiting on community ports, with NVIDIA's reference hardware reporting over 4,000 tokens/sec per GPU.
Known limitations
- No native vision, audio or video input; thinking mode is mandatory and cannot be disabled.
- DeepSWE coding benchmark score of 56.6 sits well below the field's current leaders, so this checkpoint's GPQA-level reasoning strength doesn't fully carry over to agentic coding.
- Training data cutoff date and safety system card remain unpublished as of August 2026.
Who it is for
- Agentic coding teams building self-hosted or multi-cloud pipelines who want frontier-scale reasoning without single-vendor API lock-in
- Infrastructure teams with multi-GPU or multi-node capacity able to serve a 2.4T-parameter MoE checkpoint
- Cost-sensitive teams comparing per-token price against GPT-5.6 Sol, Gemini 3.7 Flash and Claude Opus 5 for long-context reasoning workloads
- Researchers benchmarking open-weight frontier-scale MoE architectures
Supported languages
- English
- Chinese
Qwen3.8-2.4T-A95B benchmarks
Qwen3.8-2.4T-A95B against its nearest rivals
Output speed tokens per second, higher is better
Standout result
GPQA Diamond, Alibaba's reported score for its largest-ever open-weight release
Source: Hugging Face model card / NVIDIA Technical Blog, August 2026Where it sits
Priced around the middle of the 103 GA models with a published price (rank 61), and in the top third on GPQA Diamond (rank 14 of 63).
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $3.00/M | #61 / 103 | peer median $1.71/M |
| Output speedtokens/s | 38 tok/s | #54 / 57 | peer median 94 tok/s |
| GPQA Diamond% correct | 92.6% | #14 / 63 | peer median 87.8% |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Measured figures
- AA blended price
- $1.18/M
- Output speed
- 38 tok/s
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Strengths and limits
Strengths
- First Qwen-Max-class model Alibaba has open-weighted, priced below its own predecessor's rate on the official API (see the pricing FAQ for the exact numbers).
- Scores among the highest reported GPQA Diamond results in Alibaba's Qwen3.8 series to date, ahead of its immediate predecessor on graduate-level reasoning.
- Ships same-day in FP8, GGUF and NVFP4 quantized formats plus Fireworks/SGLang/vLLM/NVIDIA NIM serving support, so self-hosting doesn't require waiting on the community to catch up.
- Scores 92.6% on GPQA Diamond, Alibaba's highest-reported reasoning score in the Qwen3.8 series so far.
Show 2 more
- Prices at $2.00/$6.00 per 1M tokens on the official API, below Qwen3.7-Max's prior $2.50/$7.50 rate despite the larger parameter count.
- First Qwen-Max-class checkpoint released as open weights, with same-day FP8, GGUF and NVFP4 quantized builds.
Limits
- DeepSWE coding score of 56.6 lags the field's current top performers, so the strong GPQA Diamond result doesn't carry through to every coding benchmark.
- Text-only despite the 'Max' branding: no image, audio or video input, unlike what several third-party writeups imply.
- No published training data cutoff date or safety system card as of August 2026, which complicates enterprise compliance review.
- Custom license requires a separate commercial agreement once trailing-year revenue crosses a set threshold for hosting or AI coding/office-assistant products, unlike a standard permissive open-source release.
Qwen3.8-2.4T-A95B alternatives
The nearest AI models on Hokai for the same job.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Qwen3.8-2.4T-A95B This page | Alibaba Cloud | 1M | $2 / $6 | text → text |
 Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
| Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
 Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens. | 262K | $0.10 / $0 | text, image, audio, video, pdf → text |
Qwen3.8-2.4T-A95B specs
- Context window
- 1.0M tokens
- Max output
- 131K tokens
- Tokens per second
- 350 tok/s
- Rate limit (RPM)
- 15000
- Rate limit (TPM)
- 2,000,000
- Mixture of experts
- Yes
- Parameters
- 2.4T total, 95B active per token
- Architecture
- Mixture-of-Experts (MoE), fine-grained hybrid full + linear attention
Training data
Not disclosed. Alibaba has not published a training data cutoff date or a description of the pretraining corpus for this checkpoint as of August 2026.
Weight formats
- safetensors (bf16)
- FP8
- GGUF
- NVFP4
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| QwenCloud API | https://www.qwencloud.com | API Key |
| OpenRouter | https://openrouter.ai/qwen/qwen3.8-2.4t-a95b | API Key |
Regions
- International (Singapore)
- Mainland China (Beijing)
Prompting patterns
Show 2 prompting patterns
Set reasoning effort explicitly
Pass a reasoning_effort value of low, high, or xhigh per request
The model defaults to always-on thinking; picking the right tier avoids paying xhigh-level latency and token cost for simple requests.
Guard against image/video input at the API wrapper level
Reject or reroute requests containing image_url or file attachments before they reach qwen3.8-2.4t-a95b
The checkpoint is text-only and will not cleanly error on multimodal input; wrapping this at the client avoids hallucinated image descriptions.
Known quirks
Show 5 known quirks
Text-only despite 'Max' branding and press claims
Several third-party writeups describe the hosted Qwen3.8-Max product as accepting image and video input, but this checkpoint's own Hugging Face model card states multimodal input is not supported and thinking mode cannot be disabled. The same confusion was already documented on Qwen3.7-Max, where users expected vision support that only the separate Plus variant provides.
Workaround: Confirm modality support against the Hugging Face model card before routing image or video requests to this checkpoint. Route multimodal tasks to Qwen3-VL or a future Qwen3.8-Plus-class release instead.
Native context is 262K, not the marketed ~1M
The model's native context length is 262,144 tokens; reaching the documented ceiling of roughly 1,010,000 tokens requires the extended-context configuration, which most third-party summaries collapse into a flat '1M-token context' headline.
Workaround: Check your serving stack's rope-scaling or extended-context config before assuming the full ~1M window is active by default, especially on self-hosted vLLM or SGLang deployments.
Reported SWE-bench score is the harder 'Pro' variant, not 'Verified'
Alibaba's published 67.7% figure is on SWE-bench Pro, a newer and harder benchmark than the more commonly cited SWE-bench Verified. Comparing it directly to older SWE-bench Verified scores from other models overstates or understates the gap depending on which model it's compared against.
Workaround: When benchmarking against other models, confirm which SWE-bench variant (Verified vs Pro) their published score uses before drawing a head-to-head conclusion.
Custom license, not Apache 2.0, despite the 'open weights' framing
Coverage often calls this an open-source release, but the Hugging Face license terms impose a UI-attribution requirement above 100 million MAU and require a separate commercial license above $50 million in trailing 12-month revenue for model-hosting or AI coding/office-assistant products.
Workaround: Read the actual LICENSE file on the Hugging Face repo before assuming Apache-2.0-style unrestricted commercial use, particularly if you plan to resell inference access.
A newer hosted-only Max-0902 snapshot exists and is not this checkpoint
Alibaba shipped a post-training upgrade, Qwen3.8-Max-0902, to the hosted QwenCloud API on September 2, 2026, at unchanged pricing but with substantially better agentic-coding scores. It is API-only -- Alibaba has not open-weighted it -- and does not replace this Hugging Face checkpoint, which remains the only downloadable Max-class Qwen weights.
Workaround: If comparing benchmark numbers published after September 2026, confirm whether the source tested the open-weight Qwen3.8-2.4T-A95B checkpoint or the hosted qwen3.8-max-0902 snapshot; the two score differently on the same benchmarks.
Qwen3.8-2.4T-A95B facts
- Deployment
- API · Self-hosted · NVIDIA NIM
- Safety posture
- Balanced
- Released
- 12 Aug 2026
- Current version
- Qwen3.8-2.4T-A95B
- Family
- Qwen3.8
- Variant
- Chat
- Licence
- Custom Qwen open-weight license (not Apache 2.0): 100M+ MAU triggers a UI attribution requirement, and $50M+ trailing-12-month revenue from model-hosting or AI coding/office-assistant products requires a separate commercial license from Qwen. · Open weights
- Data retention
- Not publicly documented for this specific model as of August 2026; hosted usage falls under Alibaba Cloud's general QwenCloud/DashScope terms of service.
Safety controls
- Rlhf
- Yes
- System prompt
- Yes
- Constitutional AI
- No
Data governance
- Data residency options
- International (Singapore), Mainland China (Beijing)
Version history
Show all 5 versions
Qwen3.8-Max-0902 · 02 Sep 2026
Hosted-only post-training upgrade on the QwenCloud API: unchanged architecture and pricing, substantially improved agentic-coding scores. Not open-weighted and does not replace this checkpoint.
Qwen3.8-2.4T-A95B · 12 Aug 2026
Open-weight release of the Qwen3.8-Max checkpoint: safetensors, FP8, GGUF and NVFP4 builds on Hugging Face, Day-0 serving support from Fireworks AI, SGLang, vLLM and NVIDIA NIM/Dynamo.
Qwen3.8-Max · 03 Aug 2026
General availability on the QwenCloud API at $2.00/$6.00 per 1M tokens, succeeding Qwen3.7-Max.
Qwen3.8-Max-Preview · 19 Jul 2026
Alibaba's first public preview announcement of the 2.4T-parameter Qwen3.8 flagship.
Qwen3.7-Max · 20 May 2026
Predecessor flagship, announced at the Apsara Summit in Hangzhou; proprietary, text-only, supported autonomous coding runs up to 35 hours.
What changed
Questions about Qwen3.8-2.4T-A95B
What does Qwen3.8-2.4T-A95B cost to run via the API?
QwenCloud's international (Singapore) endpoint bills this model at $2.00 for every 1M input tokens processed and $6.00 per 1M output tokens generated, with an 8x discount on cached input at $0.25 per 1M. That undercuts the prior-generation flagship's higher per-token rate, and OpenRouter routes the same model through multiple upstream hosts at a lower blended rate, close to $1.80/$5.40 per 1M. See the price notes above for the Mainland China endpoint's regional discount.
How does Qwen3.8-2.4T-A95B perform on independent and third-party benchmarks?
Alibaba's own model card reports 67.7% on SWE-bench Pro for this checkpoint, matching the figure covered in the full review above. On the independently tracked DeepSWE coding benchmark, though, it scores well below its graduate-level reasoning result, and Artificial Analysis's Intelligence Index likewise lands on the lower end, showing the model's reasoning strength doesn't carry through to every coding or composite benchmark equally. AIME 2025, MMLU-Pro and SWE-bench Verified scores remain unpublished as of this writing.
Is Qwen3.8-2.4T-A95B open source or proprietary?
This checkpoint ships as open weights on Hugging Face in safetensors, FP8, GGUF and NVFP4 formats: the first Qwen-Max-class release Alibaba has open-weighted. The license is custom to Qwen, not a standard permissive open-source license, and layers in two size-based obligations once a product scales: a UI-attribution requirement at high user counts, and a separate commercial agreement once trailing-year revenue from model-hosting or AI coding/office-assistant products crosses a set threshold. Read the LICENSE file on the Hugging Face repo for the exact figures.
Does Qwen3.8-2.4T-A95B train on user data?
Alibaba has not published a training data cutoff date or a system card for Qwen3.8-2.4T-A95B as of August 2026. For hosted use through QwenCloud, Alibaba's general terms of service govern data handling; no zero-retention or enterprise data-processing agreement specific to this model has been publicly documented yet.
Which teams should choose Qwen3.8-2.4T-A95B, and who should look elsewhere?
Qwen3.8-2.4T-A95B suits teams running self-hosted or multi-cloud agentic coding pipelines who want frontier-scale reasoning without single-vendor API lock-in. Teams needing native image or video input should look elsewhere: the checkpoint is text-only despite the 'Max' branding. Teams needing a published safety system card for compliance review may prefer Claude Opus 5 or GPT-5.6 Sol, which both document red-teaming partners.
About Qwen3.8-2.4T-A95B
Qwen3.8-2.4T-A95B is Alibaba Cloud's largest foundation model release to date, open-weighted on Hugging Face on August 12, 2026. It is the checkpoint behind the QwenCloud-hosted Qwen3.8-Max product, GA'd on August 3, 2026. The architecture is a fine-grained mixture-of-experts transformer: 2.4 trillion total parameters, 95 billion active per token, across 512 experts (10 routed plus 1 shared per token) and 92 layers interleaving linear and full attention in a 3:1 pattern. It succeeds Qwen3.7-Max, announced at the Apsara Summit on May 20, 2026, and is the first Qwen-Max-class flagship Alibaba has open-weighted rather than kept API-only.
Read the full background
Alibaba reports 92.6% on GPQA Diamond, 67.7% on SWE-bench Pro, 86.6% on Terminal-Bench 2.1, 86.1% on OSWorld-Verified, 93.0% on PaperBench and 82.8% on IFBench; no SWE-bench Verified score has been published for this checkpoint by Alibaba, Artificial Analysis or any independent tracker as of September 2026. On the independently tracked DeepSWE 1.1 coding benchmark it scores 56.6, trailing several frontier agentic-coding competitors despite the strong GPQA result. Artificial Analysis puts its Intelligence Index at a comparatively modest 40 and its blended price at $1.18 per 1M tokens (7:2:1 cache/input/output ratio), an independent cross-check on Alibaba's own figures. AIME 2025, MMLU-Pro and ARC-AGI 2 remain unpublished, and independent verification is limited given the model's age.
Native context is 262,144 tokens, extensible to roughly 1,010,000, which is where the "1 million token context" figure in most coverage comes from. Recommended max output is 131,072 tokens for a final answer, though up to 262,144 of that budget can go to visible reasoning traces first, since thinking mode is mandatory and cannot be disabled.
Despite the "Max" branding, Qwen3.8-2.4T-A95B is text-only: its Hugging Face model card confirms multimodal input is not supported. It accepts text and tool-call input and returns text and tool-calls, with native OpenAI-compatible function calling, structured output and parallel tool calls. This mirrors Qwen3.7-Max, whose text-only "Max" flagship was often confused with the multimodal "Plus" variant; several writeups already describe Qwen3.8-Max as accepting image and video input, which this checkpoint does not. It ships configurable reasoning: low, high and xhigh effort levels per request.
QwenCloud's International (Singapore) endpoint undercuts Qwen3.7-Max's prior per-1M-token rate despite the larger size, with a steep discount on cached input (exact figures in the pricing FAQ below). The Mainland China endpoint runs an estimated 60-70% cheaper, and OpenRouter's routed price runs lower still. A 1M-in/200K-out coding-agent run costs about $3.20; a 1,000-turn support chat averaging 2K in/500 out per turn runs about $7.00.
Open weights ship on Hugging Face as safetensors, FP8, GGUF (Unsloth) and NVFP4 (RadixArk) builds. Day-0 serving came from SGLang, vLLM, NVIDIA NIM/Dynamo and Fireworks AI, alongside the official QwenCloud API and Alibaba Model Studio. NVIDIA reports over 4,000 tokens/sec per GPU and 350 tokens/sec per user on its GB300 NVL72 reference system, a figure specific to that deployment rather than any hosted API's guaranteed rate. No AWS Bedrock, Google Vertex or Azure availability is confirmed yet.
Alibaba has not published a safety system card, training data cutoff, or named red-teaming partners for this model. Its custom license, not Apache 2.0, requires products over 100 million monthly active users to display the model name in their UI, and "Model as a Service" or "AI Work Assistant" businesses over $50 million in trailing 12-month revenue to sign a separate commercial license before commercial use.
It suits teams building self-hosted or multi-cloud agentic coding and long-document pipelines who want frontier-scale reasoning without single-vendor lock-in, and who have the multi-GPU infrastructure a 2.4T-parameter checkpoint needs even at FP8 or NVFP4. Cost-sensitive teams often weigh it against DeepSeek V4, Kimi K3 and Gemini 3.7 Flash for the same long-context reasoning workloads. Teams needing native image or video understanding should use Qwen's separate VL line, or OpenAI's multimodal models, instead; teams needing a published system card for compliance may prefer GPT-5.6 Sol or Anthropic's Claude Opus 5, which document red-teaming partners this model lacks. Browse the rest of the large language model category for open- and closed-weight alternatives.
Sources and guides
Official sources
- Qwen3.8-2.4T-A95B model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Alibaba Cloud's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides that mention Qwen3.8-2.4T-A95B
Lists that include Qwen3.8-2.4T-A95B
The provider
More AI Models on Hokai
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.
Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.
JevTypeSafe's first System One Model, launched in 2026: skips text generation for typed, calibrated decisions software can act on directly.Nex-N2.6Nex AGI's hosted agentic model for computer use and coding, with image input, tool calling and about 321B parameters, served first-party on OpenRouter.